Getting Started¶

1. Set up the environment¶

make setup
conda activate hackathon

This installs Python, all scientific deps, FluidSynth (the C library), and pyfluidsynth (Python bindings) in one shot.

2. Verify audio works¶

make music

This runs the src/midi_demo.py script, which cycles through instruments and chords to ensure FluidSynth is correctly configured and communicating with your speakers.

If you hear nothing, check that your system volume is on and the correct audio driver (WASAPI is preferred) is available in Windows Settings.

3. Launch the GUI¶

# Without hardware (mock data for development):
make gui

# With BioRadio connected:
make gui-live

4. Train the classifier¶

The ML pipeline lives in src/pipeline.py. Victor's main() trains on the team's 8 gesture classes from recorded EMG data:

make train

This produces a saved model in models/classifier.pkl.

5. Perform the Ritual¶

Once your classifier is trained and the GUI is streaming to LSL, start the real-time bridge:

make ritual

This will:

1. Initialize the MIDI engine (FluidSynth + SoundFont).
2. Load models/classifier.pkl (falls back to a mock classifier if not found).
3. Search for the BioRadio LSL stream (retries up to 8 times, 1 second each).
4. Process EMG signals in 250ms windows with 50% overlap.
5. Classify gestures and trigger the MIDI engine.

You can also start the ritual from the GUI by enabling the Music Mode checkbox during acquisition. The GUI shows real-time status updates as the ritual connects.

6. Manual Integration (Alternative)¶

If you're building a custom script, connect MIDI to your own classifier loop:

from midi_engine import MidiController

controller = MidiController()
controller.start()

# Called every classification frame:
controller.on_classification(
    right_hand="palm_up_out",   # chord selection
    left_hand="fist_down_out",  # instrument selection
    amplitude=0.73              # EMG amplitude -> velocity
)

controller.stop()